Upgrade to Pro
— share decks privately, control downloads, hide ads and more …
Speaker Deck
Features
Speaker Deck
PRO
Sign in
Sign up for free
Search
Search
CVPR2019 参加速報 本会議1日目 / CVPR2019 Personal Memo: Day 1
Search
Atsushi
June 18, 2019
Technology
0
350
CVPR2019 参加速報 本会議1日目 / CVPR2019 Personal Memo: Day 1
チラシの裏チラシの裏チラシの裏チラシの裏チラシの裏チラシの裏
Atsushi
June 18, 2019
Tweet
Share
More Decks by Atsushi
See All by Atsushi
CVPR2019参加速報 本会議 3日目 / CVPR2019 Personal Memo: Day 3
atsushihashimoto
0
260
CVPR2019参加速報 本会議 2日目 / CVPR2019 Personal Memo: Day 2
atsushihashimoto
0
220
Other Decks in Technology
See All in Technology
Cloudflare WorkersがPythonに対応したので試してみた
miura55
0
190
OPENLOGI Company Profile for engineer
hr01
1
2.2k
生成AIと産業向けソフトウェアの自動生成 〜 ハノーバーメッセ2024より〜
kioto
2
420
汎用ポリシー言語Rego + OPAと認可・検証事例の紹介 / Introduction Rego & OPA for authorization and validation
mizutani
1
140
データ基盤を支える技術
chanyou0311
5
2.9k
日本が誇るイタリアのダンスミュージック!? ユーロビートって何??
minorun365
PRO
2
190
iThome2024 Wailing Wall of Enterprise Security
notsurprised
0
280
複雑なビジネスルールに挑む:正確性と効率性を両立するfp-tsのチーム活用術 / Strike a balance between correctness and efficiency with fp-ts
kakehashi
5
3.5k
YJIT Makes Rails 1.7x faster / RubyKaigi 2024
k0kubun
3
440
DevRelによる信頼構築とデータ駆動で変わるエンジニア採用 / DevRel Trust Building to Data Driven Engineering Hiring
bobtani
1
130
Password cracking: past, present, future
openwall
0
250
TypeScript の抽象構文木を用いた、数百を超える API の大規模リファクタリング戦略
yanaemon
6
1.2k
Featured
See All Featured
Git: the NoSQL Database
bkeepers
PRO
423
63k
Put a Button on it: Removing Barriers to Going Fast.
kastner
58
3.1k
Fantastic passwords and where to find them - at NoRuKo
philnash
39
2.5k
Visualization
eitanlees
137
14k
What the flash - Photography Introduction
edds
64
11k
No one is an island. Learnings from fostering a developers community.
thoeni
16
2.1k
Robots, Beer and Maslow
schacon
PRO
155
8k
The Cost Of JavaScript in 2023
addyosmani
21
4k
Producing Creativity
orderedlist
PRO
338
39k
Bash Introduction
62gerente
605
210k
Typedesign – Prime Four
hannesfritz
36
2.1k
Navigating Team Friction
lara
179
13k
Transcript
CVPR2019 ຊձٞॳ ใ
- ࡢ·Ͱͱಉ༷ɺݸਓͷϝϞΛެ։͍ͯ͠ΔΑ͏ͳܗͷͷͰ͢ɻ͋͘·Ͱ ɺͪΒ͠ͷཪతͳѻ͍Ͱ͓ئ͍͠·͢ - ࢲͷཧղͷँΓଟʑ͋Δͱࢥ͍·͢ͷͰ͝༰͍ࣻͩ͘͞ɻ ࠓճ͔Βޱ಄ൃදγϣʔτΦʔϥϧͷΈʹͳΓ·ͨ͠ɻ5ൃදx3݅ຖʹ3ͷ࣭͕ٙ͋Δܗ(ܭ18) ͰਐΜͰ͍͘ܗͰ͢ɻΈΜͳख๏ͷৄࡉ·ͰͤͣɺͪΐͬͱUXͷ͍ܗࣜͩͬͨ=ࣸਅͱͬͨΓϝϞ ͢ΔՋ͕ͳ͔ͬͨͷͰɺࠓճ͕ࣗฉ͍ͨൃදΛϙελʔͷΈʹߜͬͯྻڍɻ
ಉҰମผ࢟ͷ2ͭͷಛ͕Ұக͢Δ Α͏Siamese LossΛ͔͚ͭͭɺޓ͍Λઢ ܗม͢Δͱผ࢟ͷsegmentΛநग़Ͱ ͖ΔΑ͏ʹֶशˠະମͰCoSeg͕Ͱ ͖ΔΑ͏ʹͳΔɻ
؆୯ͳαϯϓϧ͔ΒঃʑʹDA͢Δ ͱͪΐͬͱਫ਼͕͋Δɻ SOTAʹશવಧ͍͍ͯͳ͍Α͏ʹ Έ͑Δɻ
γʔϯ͔ΒActorݕग़ͯ͠ɺ Actor͝ͱʹ࣍ʹͲ͏ͳΔ͔Λ ༧ଌˠMessage PassingΛ܁Γฦ͢ ͜ΕʹΑΓಈ࡞༧ଌਫ਼Λ্ɻ
͋ͱͰಡΉɻ ಛྔΛ͢Δɻ ஶऀෆࡏɻ
Multi-label classificationͰɺ֤ଐੑͷ αϯϓϧͷInbalanceΛௐ͢ΔLossΛఏҊ
Region Proposal Network͕Γग़͠ ֤ۣͨܗʹରͯ͠Adv. trainingͰ Domain AdaptationΛ͔͚͍ͯΔͬΆ ͍
λΠτϧΛΈͯԿ͔ͱࢥͬ ͕ͨɺࠨଆͷྻͷઃఆ ΛݟΔʹ Open-setͷख๏ͬΆ ͍ɻ
ΊͬͪΌݟΒΕͯͨ
୯Ұը૾ͷΈ͔Βɺ UnsupervisedͰ ಈ͍͍ͯΔମΛݕग़͢Δख๏ɻ Ͳ͏ɺγʔϯͷ͏ͪɺ͔̍ͭ͠ ग़ྗ͠ͳͦ͞͏ͩͬͨͷͰ ৄࡉεΩοϓ
γΣΠϓಛͱ࢟ಛΛ ൈ͖ग़ͯ͠ɺܗͷ ʮΛணͨਓʯϞσϧΛ ೖྗ͢Δͱɺணͤସ͕͑Ͱ͖ Δʁ
2Dͷ࢟ਪఆx2ຕʹରͯ͠ɺ epipolarͰ3࣍ݩΛ෮ݩ͠ɺ1ຕ͔Β 3Dͷ࢟Λग़ྗͰ͖ΔNNΛผ్ ֶश=self supervised.
LSTMʹ͔͚ΔͷͰͳ͘ɺ ஷΊ·͍ͬͯ͑͘ͱؔʹ͔͚ Δʢؔͷৄࡉෆ໌ʣ Epic-Kitchenͷ݁ՌΛݟΔͱ ͘͢͝ޮ͘ɺͱ͍͏ҹ ͋·Γͳ͍͔ɻ ͬͱଞʹํ๏͕͋Γͦ͏ ڞஶऀ߽՚ɻ
ະདྷ༧ଌͷͰ͖ͳ͍ͱ͜Ζ= EventͷΕͱ͢Δɻ લ͔Βࢲ͕͍͍ͬͯΔख๏ɻ ·͊ɺ͋ΔఔͰ͖ΔΑ͏ʹ ࢥ͑Δ͕ͦͷઌ͕ͩͱࢥ ͏ɻ
How to Do 100MΈ͍ͨͳ ͷɻ.͚ͩͲ
LSTMͰ1ϑϨʔϜग़ྗɺ ͦͷϑϨʔϜΛઌ಄ͱͯ͠ɺ ٯ͖ͷLSTMΛֶश͠ɺ ઌ಄ϑϨʔϜΛੜ͢Δ Cycle GANɻͳ͔ͥMSE͕ ͕͍͋ͬͯΔɻ ࣭͕ͨ͠ɺߟ͍͑ͯͳ͔ͬ ͨΒ͘͠ɺ͔֬ʹͳΜͰͩΖ ͏ͱஶऀ͕ΜͰ͍ͨ...
Self-supervised Learning͕ AlexNetҙ֎ͰͪΌΜͱಈ ͘ͷ͔Λௐͨจɻ Take Home Message͕وॏ
Text > Image > Text ͷCycleͷ ΈGANΛద༻ͯ͠ɺText-to- imageͷม࣌ͷใଛࣦΛݮ Βͨ͠ɻ
Yale SongͷൃදɻDiversity Lossͱ͔ࣅͨΑ ͏ͳͷ͋ΔͣɺΈ͍ͨͳ͜ͱΛ͍͍ͬͯ ͨɻ
Ranjay Krishnaͷൃදɻͪ͜Β Θ͔Γ͍͢ɻ ࣭จΛੜ͢ΔͷΛֶशɻ 1. ը૾ͱਖ਼ղΛೖྗͱ࣭ͯ͠จ Λੜɻ 2. (ೖྗʹਖ਼ղ͕͋Δͱ࣮༻ੑ͕ͳ ͍ͷͰ)ਖ਼ղˠਖ਼ղΧςΰϦʹೖΕ
ସ͑ͨωοτϫʔΫͰɺಉ͡ಛ ͕ग़ྗ͞ΕΔΑ͏ʹֶशɻ
Knowledge GraphΛೖΕͯ HOIͷݕग़Λݡ͘͢Δख๏
3D point cloud͔Βͷ ͷඪຊநग़Λϝλֶश
͋Μ·Γ৽نੑ͕Α͘Θ ͔Βͳ͍??
JigsawΛͬͯෳυϝ Πϯͷը૾Ͱֶश͢Δ ͱɺDomain Generalization(PACS)Ͱ SOTA͕ͰΔ... ҰԠɺҰ͚ͭͩτϦοΫ ͕͋ͬͯɺग़ྗϕΫτϧ ͷΤϯτϩϐʔΛ͘͢ ΔΑ͏ʹɺͭ·Γɺৗʹ ֬৴Λͬͯ͑ΔΑ͏
ͳLossՃ͍ͯ͠Δͱ ͷ͜ͱɻ
Target Domain͕ɺ࣮ࡍʹ(ະ ͷ)ෳυϝΠϯͷू߹ʹͳ ͍ͬͯΔ߹Λߟ͑ɺTarget sub-domainͷਪఆΛ(Ϋϥελ ϦϯάͰΓͳ͕Β)Α͋͘Δ UDAΛ͢Δख๏ɻ ୯ʹಛྔΛΫϥελϦϯά͢ ΔͱΧςΰϦ͝ͱͷΫϥελ͕ Ͱ͖Δةݥ͕ߴ͍ͷͰɺݩͷը
૾ͱɺಛྔΛ߹Θͤͨͷʹ ͍ͨͯ͠ΫϥελϦϯάΛ͢Δ ͱͷ͜ͱɻ
ֶशσʔλ͕Ұ༷ʹͳΔΑ͏ ͳAdversarial TrainingΛͯ͠ɺҰ ༷ʹΒͳ͍ͷΛɺHard Negativeͱͯ͠ݕग़͍ͯ͠Δɻ
ମࣝผͱಈ࡞ࣝผΛֶशͤ͞Δ͜ͱͰɺ ମۣܗΛݕग़͢ΔωοτϫʔΫΛֶशɻ ͜Εͦ͜ɺFirst Person VisionͰطʹ͋Δɻ
None
AEͷLatent Featureʹରͯ͠ ಛ্ۭؒͷڑʹج͍ͮͯ ҟৗݕ͢Δͱɺ্ख͘Open- Set͕ղ͚Δɺͱ͍͏ɻ ౦େͷݚڀɻ
ैདྷͷSpectral Net͕୯ʹ Siamese NetworkͰϓϨτ Ϩʔχϯά͍ͯͨ͠෦Λ վળɻ
ࣗಈ༁ͷSOTAʹͳͬͨͷͱಉ͡Ͱɺ image - text ͷCycle-GANΛֶशɻ
Star-GANͰม1:1ɻ͜ͷख๏ ಉ͡ରͷෳυϝΠϯͷσʔλΛೖྗͱ ͯ͠ɺλʔήοτυϝΠϯͷσʔλΛੜ Ͱ͖ΔΑ͏ʹCycle Consistency LossΛ গ͠ɻ
Conditional GANͳͲͷ conditionʹϊΠζ͕͋Δͱ͠ ͯɺͲ͏Fix͢Δ͔ɻ ֶश࣌ʹಉׂ͡߹ͷϊΠζΛࡌͤ Δɻ ஶऀʹΑΕɺϓϥϚΠ0.2͘Β ͍ͷޡࠩ͑Δɻಉ༷ʹϥϕ ϧʹϊΠζͷͳ͍σʔλʹରͯ͠ ख๏Λద༻ͯͦ͜͠·ͰѱӨ
ڹͳ͍ͱͷ͜ͱɻ จதʹσʔλ͋Γ??
None